Gemini 3 Pro tops ARC‑AGI‑2 with 31%

In November 2025, Gemini 3 Pro became the first AI model to surpass the 23 percent threshold on the ARC-AGI-2 benchmark, achieving an actual score of 31 percent according to reports shared by Ethan Mollick. This milestone highlights the continued leadership of closed-weight frontier models while underscoring an 8-to-12-month performance gap that persists with open-weight alternatives.

Key Takeaways

• Closed models like Gemini 3 Pro maintain a clear lead on complex reasoning benchmarks such as ARC-AGI-2, creating strategic advantages for enterprises needing reliable high-level cognition.
• Open-weight models including GLM-5.2 are closing the gap on select tasks but exhibit jagged performance profiles that require careful evaluation before deployment.
• Businesses can monetize these developments by integrating hybrid AI stacks that combine proprietary reasoning engines with cost-effective open models for specialized workflows.

Deep Dive into ARC-AGI-2 Performance

The ARC-AGI-2 benchmark measures abstract reasoning capabilities that go beyond standard language tasks, testing a model's ability to solve novel puzzles with minimal examples. Gemini 3 Pro's 31 percent score in late 2025 marked a significant leap, demonstrating superior generalization compared to earlier systems. Open-weight contender GLM-5.2 from Zai_org reached 22.8 percent on ARC-AGI-2 and 77 percent on ARC-AGI-1 at low cost, performing comparably to GPT-5.4 and GPT-5.5 under low reasoning effort as noted by the ARC Prize organization.

Closed versus Open Weights Dynamics

Performance remains jagged across both categories, with models excelling in certain abstraction patterns while struggling in others. This unevenness stems from differences in training data scale, alignment techniques, and architectural choices that favor specific cognitive domains. The persistent 8-to-12-month lag suggests closed labs continue to benefit from proprietary data advantages and larger compute budgets.

Business Impact and Opportunities

Enterprises focused on automation of complex decision-making can leverage closed models for mission-critical applications where accuracy exceeds 30 percent on reasoning tests, while routing simpler queries to open models to reduce inference costs. Monetization strategies include developing vertical solutions in logistics, scientific research, and software engineering that capitalize on the strengths of each model class. Implementation challenges such as integration latency and output consistency can be addressed through ensemble routing layers and continuous benchmarking pipelines.

Future Outlook

Industry analysts predict the gap will narrow further as open-weight developers gain access to larger synthetic datasets and efficient fine-tuning methods. Competitive landscapes will shift toward companies that master model orchestration rather than raw capability alone. Regulatory considerations around transparency and ethical use of reasoning systems will grow in importance, encouraging best practices such as third-party verification of benchmark claims and bias audits on abstraction tasks.

Frequently Asked Questions

What does the 31 percent ARC-AGI-2 score mean for Gemini 3 Pro?

It represents the first time any model crossed the 23 percent mark, signaling meaningful progress in abstract reasoning that can translate to better performance on novel business problems.

How does GLM-5.2 compare to closed models like GPT-5 series?

GLM-5.2 achieves 22.8 percent on ARC-AGI-2 and matches certain GPT-5 variants at low reasoning effort, offering a viable open alternative for cost-sensitive deployments.

Will the closed-open model gap close soon?

Current evidence points to an ongoing 8-to-12-month advantage for closed systems, though jagged capabilities may allow open models to lead in niche applications earlier.

What business opportunities arise from jagged model performance?

Companies can create hybrid platforms that route tasks dynamically, optimizing both accuracy and cost while building defensible intellectual property around orchestration frameworks.